计算成像是在光学成像系统中耦合了计算模型及其解耦计算方法的先进成像技术，能够突破传统成像无法逾越的物理极限，实现从宏观到微观场景的高维光信息高效采集。动态场景的高分辨率快速成像对揭示自然现象和刻画真实世界具有广泛的应用价值。然而，受限于物理原理和观测系统的局限，现有的高分辨快速计算成像技术面临诸多挑战。其中，基于单像素探测器的计算成像系统存在分辨率与成像速度的矛盾，需要设计新的压缩采样解耦算法并优化成像系统才能解决。光学层析是脑成像和肿瘤观测等生命科学关键问题的重要需求，基于相机的宽场层析成像速度快，但容易受散射和光学畸变影响导致分辨率和层析能力下降，观测活体生物还面临运动畸变和低信噪比的挑战。着眼于高分辨率快速成像的需求，本文利用计算成像的原理，通过设计新型的编码耦合成像系统以及解耦重建算法，突破现有光学系统成像能力的局限，为高分辨率单快速像素成像和高分辨率快速宽场层析成像提供了新的途径。本文的主要创新点有：1. 针对单像素成像模型中的空间分辨率与成像速度的矛盾，提出了一种内容自 适应的高维光信息压缩解耦算法，充分利用场景时间维度的连续性和局域的稀疏性差异优化结果，提升视频重建质量超过5dB。针对特定的应用场景，提出了一种基于深度学习的计算成像架构，通过神经网络联合优化计算采集 的成像过程和计算解耦的重构算法，实现了30倍成像速度提升，为单像素技术的实用化提供了解决方案。2. 针对活体环境下宽场层析显微成像受光学畸变导致分辨率下降的难题，提出了一种自适应光学校正像差的层析结构光成像方法，并针对活体生物样本的低信噪比和运动导致的重构伪值，提出了一种优化高频分量和运动补偿的结构光重构算法。研制的成像仪器在多种活体模式生物上实现了衍射极限分辨率的活体结构和功能成像，为大范围神经活动活体观测提供有效途径。3. 针对不透明生物组织内部的宽场成像因样本散射导致分辨率下降的问题，提出了双光子激发的高分辨率宽场层析成像方法，构建了线扫描结构光时空聚 焦计算显微装置，通过高速强度调制实现低成本且灵活可调的结构光照，实现分辨率近2倍的提升，有效消除了生物组织散射导致的背景噪声，为深入 散射生物体内部的高分辨率三维层析观测提供了工具。
Computational imaging, by incorporating computing into the optical imaging process, breaks the limit of conventional optical imaging and effectively achieves high-dimensional optical information acquisition for applications ranging from macro-scale to micro-scale. High-resolution imaging at high-speed is essential and promising for revealing the nature phenomena and describing how the real world works. Current high-resolution dynamic imaging techniques are facing many challenges, limited by both the physical principle and the hardware system. For example, the single-pixel detector based imaging system suffers from the trade-off between imaging speed and spatial resolution. To overcome this limitation, we must develop efficient compressed reconstruction methods together with advanced coded acquisition system. Regarding to microscopic imaging, optical sectioning is a demand for life science applications such as brain imaging and tumor observation. Fast optical sectioning techniques commonly work in a widefield configuration, the resolution and optical sectioning of which tends to degrade due to sample induced scattering and optical aberration. Sample moving and low signal-to-noise ratio bring extra chanllenges to in vivo imaging.Targeting at the demand of high-resolution fast imaging in macro- and micro-scale, this dissertation explores the idea of computational imaging by developing encoded imaging system and reconstruction algorithm to break the limitation of conventional optical imaging system. Specifically, we provide novel solutions for high-resolution single-pixel dynamic imaging and high-resolution fast optical sectioning microscopy. The main contributions of this dissertation are:1. Aiming at the trade-off between spatial resolution and imaging speed in a single-pixel model, we firstly propose a content-adaptive decoding algorithm that fully utilizes the continuity along temporal dimension and the local sparsity discrepancy, and achieves image quality improvement by 5 dB. Moreover, we propose a deep learning based computational imaging scheme, which jointly optimizes the encoded imaging process and the decoding algorithm within one end-to-end neural network. The deep learning approach enhances the imaging speed by 30 times, thus paves the way for practical usage of single-pixel imaging.2. Aiming at the resolution decrease due to optical aberration in widefield optical sectioning microscopic imaging in vivo, we propose the aberration corrected optical sectioning structured illumination microscopy with adaptive optics. We also propose a high-frequency refined and motion corrected structured illumination reconstruction algorithm, to cope with the reconstruction artifacts induced by noisy fluorescence signal and sample moving. The imaging system is demonstrated on various biological models for diffraction-limited structual and functional imaging in vivo, providing a promising solution for recording neural activity at large scale.3. Aiming at the scattering induced resolution decrease for widefield optical sectioning in nontransparent biological sample, we propose a two-photon excited high-resolution widefield optical sectioning imaging technique. We build a line-scanning temporal focusing computational imaging setup with structured illumination and implement low-cost and flexible structured illumination via rapid intensity modulation of the laser. The imaging system achieves nearly twofold resolution enhancement and effectively eliminate the sample induced scattering background signal. This work provides a tool for three-dimensional sectioning imaging deep into the scattering tissue.
